This is likely related to the point I made last beta, which is that the 2x and 4x upscale models are different and don’t produce the same results. Given the 4x came first, it may be the 2x is headed in that direction - but I really hope not as the 2x is the best version of the model I have sen so far, with the 4x much worse.
To give you an example using this current Beta, this is a comparison slider between 4x images - The original non-upscaled input, Iris V1 Auto 2x, Iris V1 Relative to Auto 2x and Iris V1 4x Upscale Relative to Auto (same settings on relative).
If you compare the Iris Relative with Iris Relative 4x you will see what a whopping difference there is between the two - and not in a good way. Iris V1 Comparison - Imgsli
On a separate note, the text enhancement issue is one of those likely unsolvable issues. If you have text in the image that is so blurred that you cannot deduce what it says, the decision needs to be made on what to do with it in an upscaled image.
Usually if the text is that bad, then its either the worst footage imaginable, something the others screenshots don’t imply, or its so small that in the original you would not have been expected to actually be able to read it. That is, the director etc was providing one of those “filler” articles where the headline is all thats important etc.
These will never be able to be enhanced properly. Based on the output, i suspect whats happening is that because it cannot read it, its replacing it with a kind of pseudo text - something that looks like text but isn’t. This would be something you do to disguise the issue.
The best example is when restoring artwork and you have to paint in the missing sections. The restorer’s job is to make the image “whole” but not under scrutiny. If you stand really close, you can find the problems, but when standing back, the image needs to look natural.
I cannot see the frame as a whole, or how noticeable your specific example is when looking at the whole frame. When upscaling the 4k its probably unavoidable to be seen - and I am not saying its doing a good job - but I wonder if what happening is its attempt to disguise the readable text in a way that you don’t get drawn to noticing something is very wrong, by replacing it with something that with only brief exposure when watching as a video, you may miss noticing its gibberish as on quick glance, on background text, it “looks” like text.
This is a screenshot of your image when bundled into this post - IE without me clicking on it to magnify. It “looks” like text and if it was not showcased, or had its attention drawn to it and only appeared briefly, you might miss completely its gibberish - and I wonder if that was the intention to it.
The rest of this message is general feedback noting that I refer always to 2x upscale here as I treat the 4x as not really usable currently.
- Comparisons between Iris Auto and Iris relative in the above show the incompleteness in detecting and sharpening the image properly. The Auto is not bad - but Auto on its own still remains the worst way to upscale anything I have tested. You must adjust those settings if you want a decent output.
- The large checkerboard pattern appears to be gone so far - at least way less noticeable if it is as I haven’t seen it yet, so that is an improvement.
- Exporting images still isn’t accurate. On this test I selected frames 58900 to 59000 and the frames export started at 58899 and finishes at 58998 meaning that of the exactly 101 frames I asked for, I got 100, with the wrong start and finish frame, so both dropped a frame, and the wrong frame sequence. I personally have no idea how exporting frames went from perfect with never an issue in V2 of Topaz, to months and months of endless problems with frames in V3, but honestly, I think its beyond a reasonable time frame to fix. I am at a loss as to how an entirely new model can be developed and tested, but if I ask for this specific extract, I cannot get it.
- Export not Export As, allows for export of frame sequences with no name and thus auto defaults to the frame number. This is great, but Export As still requires something to be put into the field on export everytime. The old version used to allow this if required, but if left blank would just export the frame numbers as default. It is a small thing, but it bugs me no end, especially when testing frequent small samples.
- I thought I would experiment with importing the original VOB to see how it would handle no pre-processing at all. When doing so, the Preview button is greyed out and won’t let me preview at all, but I can scroll the time bar and see the whole video and trim normally. Having said that:
– It won’t trim
– It jumps constantly on preview frame numbers and they are always incorrect.
– Basically it seems to have no idea how to handle the file - but gives no errors as such, just a warped user experience.
– It will not let me edit the input fps, which is incorrect for the file.
I have not seen a video application have this much difficulty understanding a vob file, and all my other applications open it with no issue.
– I attempted to export a clip of it saved directly to mkv using a different application, and this opens and seems fine with frame numbers, but it has lost its aspect ratio information and now displays wrong in Topaz.
Just a reminder that eventually I really suggest you decide at the back end to include the ability to import from Avisynth files again. I used to be able to import the vob files with little to no processing by using the d2v file assessment with it - but attempting to import the d2v file does nothing, so I assume you don’t support that anymore, and without a frame provider like Avisynth, I cannot translate that import either.
I bring this up because your goal is apparently to make it as user friendly as possible to use and upgrade your old media and home movies - but you made V3 of the program significantly harder to actually import those some media files. I have to jump through hoops that the average person doesn’t want to do just to import the same files and media V2 handled with no issue.
It is a bit of a paradox to me as to why you would intentionally make it worse without a solution to fix it in application. If I have to pre-process the files before I can feed them into Topaz, and V3 makes that harder than any previous version, then no lay person is going to be happy trying to use this application.
key in this beta - the seek bar skips a small amount of the “way” sometimes when I’m going to the next frame, and sometimes - a bigger amount. It’s totally fine, if the video has a variable frame rate.
